Taste in the age of AI and LLMs
As large language models make competent but generic output cheap to produce, commenters argue that the remaining value in software and products lies in human judgment: choosing what to build, how it should feel, and when to reject “slop.” Many are skeptical that “taste” alone is a durable moat, pointing instead to effort, proprietary data, distribution, and clear product vision as equally or more important advantages. There is broad agreement, however, that AI amplifies existing tendencies toward bland, average work and that people who can specify, critique, and refine with precision will get far more out of these tools than those who treat them as push‑button replacements for thinking.
Overall reaction to the article
- Many commenters found the piece generic, formulaic, and likely AI-generated: short declarative sentences, heavy use of bullets and subheadings, vague abstractions, and lack of concrete examples or personal anecdotes.
- Some called it ironic that an essay claiming “taste is the moat” reads like AI “slop,” exhibiting the same “empty specificity, borrowed tone, and fake confidence” it criticizes.
- A few suggested it might itself be part of a “train your taste” loop, using Hacker News as a feedback source.
Is “taste” really a moat?
- One camp agrees that judgment/taste matters more as AI makes mediocre output cheap; what differentiates people is what they choose to build, how they cut scope, and how clearly they can critique work.
- Others argue “taste” is overhyped: it’s fuzzy, varies by audience, and can be approximated by data, A/B testing, or scaled models, so it’s not a durable moat.
- Several emphasize that effort, execution speed, distribution, proprietary data, and real-world constraints still matter at least as much as taste.
What is “taste” in this context?
- Competing definitions:
- Product/PM taste: clear vision of what to build, what to reject, and how features fit together.
- Engineering taste: coherent abstractions, consistency, idiomatic patterns, and a “north star” for a codebase.
- Aesthetic taste: style, fashion, and signaling (with jokes about tech uniforms and poor tech “vibes”).
- Some note you cannot just have taste; you must also exert effort and gain skill, or you get informed complaint without good work.
AI, coding agents, and “perfect code”
- There’s active discussion around “agentic coding”:
- One approach: define in detail what “good/perfect code” means in your codebase and use LLMs under strict guidelines to raise quality and consistency.
- Critics counter that specification, review, and evolving requirements are still hard, and that “perfect” is ill-defined and often irrelevant to business success.
- Several mention “comprehension debt” and AI-created big balls of mud: AI can rapidly generate tangled codebases whose intent even AI later struggles to untangle.
- Complex domains (e.g., GPU kernels, legacy systems, security-sensitive or obscure integrations) are cited where current models still struggle badly.
Broader impacts and concerns
- Some see everyone becoming more like investors: the scarce skill shifts from doing to making good bets and allocating effort.
- Others worry about an “ocean of crap”: AI floods content and internal docs, making it harder for high-taste work to be discovered or appreciated.
- Multiple comments highlight that aligning software with messy human needs and institutions will keep human judgment critical, regardless of AI progress.